TL;DR & Quick Summary
Seasonal demand repeats, so the most important step is to use a method that knows about seasons at all. We tested five methods on 100 simulated seasonal items over a 26-week horizon. The code is plain Python with no libraries, so you can run it anywhere.
| Method | WAPE (all items) | Bias |
|---|---|---|
| 13-week moving average (ignores seasonality) | 55.3% | −7.5% |
| Seasonal naive (same week last year) | 18.2% | −9.8% |
| Seasonal naive + drift | 15.8% | −1.2% |
| Holt-Winters additive (fixed settings) | 24.4% | −1.8% |
| Holt-Winters multiplicative (fixed settings) | 21.4% | +7.2% |
Ignoring seasonality roughly tripled the error.
The simplest seasonal method won. Last year's weeks, adjusted for growth, beat Holt-Winters with fixed settings.
Multiplicative fits growing items better but overshot them by about 10%.
Key Takeaway: Start with seasonal naive plus drift as your benchmark. Keep a more complex method only if it beats that on your own data, on both error and bias.
Get Started: Want your seasonal items forecast properly? Schedule a strategy call with Cogniq AI or see our predictive analytics services.
The Methods
Seasonal naive forecasts each future week with the same week from last year. It copies the seasonal shape perfectly but ignores growth.
Seasonal naive with drift adds the average change between the last two years, so a product growing 10% a year is forecast 10% higher.
Holt-Winters (triple exponential smoothing) keeps three running estimates, the level, the trend and the seasonal pattern, and updates each one every week. It comes in two forms:
- Additive: seasonal swings stay the same size in units as sales grow
- Multiplicative: seasonal swings scale with sales, so a 30% peak stays 30% as volume doubles
The standard reference for all of these is Hyndman and Athanasopoulos's Forecasting: Principles and Practice.
The Code
All functions take a list of weekly history and a horizon h, and return a list of forecasts. m=52 is the season length in weeks.
def seasonal_naive(history, h, m=52):
last = history[-m:]
return [last[i % m] for i in range(h)]
def drift_seasonal_naive(history, h, m=52):
shift = (sum(history[-m:]) - sum(history[-2 * m:-m])) / m
last = history[-m:]
return [last[i % m] + shift for i in range(h)]
def holt_winters_additive(history, h, m=52, alpha=0.2, beta=0.05, gamma=0.2):
if len(history) < 2 * m:
raise ValueError("need at least two full seasons")
s1, s2 = history[:m], history[m:2 * m]
level = sum(s1) / m
trend = (sum(s2) - sum(s1)) / (m * m)
season = [x - level for x in s1]
for t in range(m, len(history)):
y = history[t]
prev_level = level
level = alpha * (y - season[t % m]) + (1 - alpha) * (level + trend)
trend = beta * (level - prev_level) + (1 - beta) * trend
season[t % m] = gamma * (y - level) + (1 - gamma) * season[t % m]
n = len(history)
return [level + (i + 1) * trend + season[(n + i) % m] for i in range(h)]
The multiplicative version replaces subtraction with division in the seasonal parts:
def holt_winters_multiplicative(history, h, m=52, alpha=0.2, beta=0.05, gamma=0.2):
if len(history) < 2 * m:
raise ValueError("need at least two full seasons")
s1, s2 = history[:m], history[m:2 * m]
level = sum(s1) / m
trend = (sum(s2) - sum(s1)) / (m * m)
season = [x / level for x in s1]
for t in range(m, len(history)):
y = history[t]
prev_level = level
level = alpha * (y / season[t % m]) + (1 - alpha) * (level + trend)
trend = beta * (level - prev_level) + (1 - beta) * trend
season[t % m] = gamma * (y / level) + (1 - gamma) * season[t % m]
n = len(history)
return [(level + (i + 1) * trend) * season[(n + i) % m] for i in range(h)]
The three settings control how quickly each estimate reacts: alpha for the level, beta for the trend and gamma for the seasonal pattern. We used fixed values of 0.2, 0.05 and 0.2, which is roughly how many spreadsheets and basic tools run Holt-Winters. Forecasting libraries such as statsmodels can fit these settings to each item automatically, which usually improves the results.
What We Tested
We simulated 100 items over three years of weekly sales with a fixed random seed:
- 40 stable items: a yearly pattern with no growth
- 40 growing items: the same pattern, with sales growing steadily, so seasonal swings grow too
- 20 shifting items: the seasonal peak moved a few weeks earlier in the final year
We held out the last 26 weeks, forecast them with each method, and measured WAPE (total error as a share of demand) and bias (whether the forecast ran high or low). Our guide to forecast accuracy metrics in Python explains both.
These are simulated results. Real items are messier, but the way each method behaves carries over.
Results by Type of Item
| Method | Stable | Growing | Shifting season |
|---|---|---|---|
| 13-week moving average | 55.3% | 54.3% | 58.2% |
| Seasonal naive | 13.6% | 19.2% | 23.3% |
| Seasonal naive + drift | 13.8% | 13.1% | 27.9% |
| Holt-Winters additive | 14.5% | 25.6% | 38.0% |
| Holt-Winters multiplicative | 14.0% | 17.2% | 47.6% |
Figures are WAPE. Lower is better.
What the Results Show
1. Ignoring seasonality is the costliest mistake. A moving average forecasts a flat line, so it misses every peak and trough. It had about three times the error of any seasonal method. If your current tool forecasts seasonal items with a moving average, fixing that is the biggest single improvement available.
2. On stable items, every seasonal method did about equally well (13.6–14.5%). When nothing is changing, last year's pattern is hard to beat.
3. Growth needs handling. Plain seasonal naive ran 17% low on growing items, because last year's sales were lower. Adding drift cut both the error (19.2% to 13.1%) and the bias (−17.2% to −0.8%). Multiplicative Holt-Winters also handled growth better than additive (25.6% to 17.2%), as expected when swings grow with sales, but it overshot by about 10%.
4. A shifting season hurts everything, and slow-adapting methods most of all. When the peak moved, Holt-Winters kept forecasting the old pattern for longer, and its error rose to 38–48%. Plain seasonal naive coped best because it simply copies the most recent year.
5. More complex isn't automatically better. With fixed settings, Holt-Winters didn't beat the simplest seasonal method overall. Fitting its settings per item would likely close some of the gap. But the comparison shows why a simple seasonal benchmark must always be in the test.
Choosing Holt-Winters Settings
If you use Holt-Winters, its three settings matter more than which version you pick. Each one is a number between 0 and 1 that controls how quickly an estimate reacts:
- alpha (level): higher reacts faster to changes in overall sales, but also chases noise
- beta (trend): keep this low. A trend that reacts too quickly turns a good month into runaway forecasts.
- gamma (season): higher adapts faster when the seasonal pattern moves, which would have helped our shifting items
Fixed settings, as we used, are a reasonable starting point but rarely the best for every item. The better approach is to fit them per item on past data, either by trying a grid of values against held-out weeks or by letting a forecasting library estimate them. Whichever you do, judge the result on held-out data, not on how well it fits the history it was trained on.
Check the Forecast Before the Peak
Seasonal forecasts matter most just before the busy period, when ordering decisions are made. A few weeks before each peak:
- Compare this year's run-up with last year's. If the early weeks are running well above or below, adjust before the peak rather than after.
- Look at the items with the largest forecast errors last season and check whether anything has changed: new competitors, price changes, a new channel.
- Confirm event dates. Moving holidays and promotions are the most common reason a good seasonal model gets the timing wrong.
How to Choose in Practice
| Situation | Start with |
|---|---|
| Stable seasonal items | Seasonal naive |
| Growing or declining items | Seasonal naive + drift, or multiplicative Holt-Winters with fitted settings |
| Seasonal swings that grow with sales | Multiplicative methods |
| Seasons that move (holidays, weather, campaigns) | Shorter-memory methods, plus known event dates as inputs |
| Fewer than two years of history | Borrow the pattern from similar items or the category |
Whatever you choose, run it against seasonal naive plus drift on held-out history and keep it only if it wins on both WAPE and bias.
Moving Holidays and Events
A fixed 52-week season assumes peaks fall in the same week each year. Many don't. Easter, Ramadan, Diwali and Chinese New Year move, and so do promotions and school holidays. Weekly seasonal methods smear these peaks across neighbouring weeks.
The practical fix is to record event dates and treat them separately: remove the event uplift before fitting the seasonal pattern, forecast the baseline, then add the uplift back in the right week. That's also why the shifting-season items in our test were hardest for every method.
From Forecast to Stock
A seasonal forecast is only useful if it changes what you order and when:
- Order ahead of the peak, allowing for supplier lead times. A forecast that sees the peak but arrives after the order deadline doesn't help.
- Set safety stock from forecast error, which is usually larger in peak weeks. See safety stock and reorder point in Python.
- Plan the run-down after the peak, so you aren't left with excess stock.
- Classify first. Our ABC-XYZ analysis in Python shows that seasonal items are often wrongly labelled unpredictable.
If your items sell irregularly rather than seasonally, use Croston, SBA and TSB instead.
Common Mistakes
- Using a moving average for seasonal items. It's the default in many spreadsheets and the most expensive mistake in our test.
- Too little history. One season can't separate a pattern from a one-off. See how much data demand forecasting needs.
- Ignoring growth. Last year's peak, unadjusted, under-forecasts a growing product every peak.
- Choosing additive when swings grow with sales, or multiplicative when they don't.
- Judging on error alone. The multiplicative method looked good on growing items until you checked its bias.
- Leaving stockouts in the history. A peak capped by a stockout teaches every method to under-forecast next year. See how to estimate lost sales from stockouts.
Conclusion
For seasonal demand, the biggest gain comes from using any seasonal method at all. In our test that cut error from 55% to under 20%. Among seasonal methods, last year's weeks adjusted for growth were the strongest overall, while Holt-Winters with fixed settings didn't justify its complexity. Use the simple method as your benchmark, try fitted Holt-Winters against it, and judge on both error and bias.
Schedule a strategy call with Cogniq AI to forecast your seasonal items properly, or explore our predictive analytics services.



